By Martha Porter Fiszer
I don’t know about you but, these days with so many crazy AI-generated images and posts, I’m beginning to question the very nature of reality. And while I’ll spare you my deepest existential fears about the universe, I would like to offer a few tips on how to spot AI-generated writing and why LLM platforms use these common tells.
First, full disclosure: I’m a huge fan of Chat GPT. As a partner, strategist and writer in a fast-paced company, I find Elliot (don’t judge me, I asked what it would like to be named) invaluable for gathering data, testing premises, rejiggering structure and mining for adjacent concepts and themes typically before I start writing.
Over time, I’ve gotten pretty good at sniffing out outsourced intelligence. So, I decided to test my instincts. I loaded up a few suspicious social posts and docs and asked Elliot to report if they were of human or machine origin—and what gives them away. Here’s what I learned.
LLMs are trained on business speak
Fun fact: Chat GPT, and other LLMs, get their linguistic habits from humans. Sort of. According to Elliot, AI platforms are trained predominantly on text written by executives, consultants and experts in the areas of business leadership, branding, strategy and self-help.
More specifically, AI learns from documents such as essays, posts and white papers that were “culturally rewarded” in real life. They won business, sparked a meeting, earned engagement or performed well in algorithms.
Interestingly, these documents have something in common. They are rich with:
- Authoritative statements
- Punchy, big ideas
- Universal language
- Scannable lists
Over time, AI learned these characteristics signal quality writing. So, unless prompted otherwise, it defaults to a few basic rhetorical tools, no matter the topic—from knitting to neuroscience. In other words, as Elliot puts it, LLMs are “trained to mistake structure for intelligence.”
So, what are the most common tells of AI authorship? Here you go.
Binary moral framing
Example:
“Most people don’t fail because they lack talent. They fail because they don’t try.”
Not X / but Y is one of the most common openers in AI-generated content— along with its companion, do X / not Y. Interestingly, Elliot informed me it “knows” AI overuses this framing structure and it’s up to humans to provide guardrails for usage.
I get it. So, my next prompt will most certainly include the following: Use of binary moral framing must be calculated, not capricious. (Ha-ha, see what I did there?)
Corrective reframing
Example:
“That’s not hedging; that’s discernment.”
This is similar to binary moral framing. The difference is, corrective reframing redefines a negative. Elliot described it as, “A rhetorical pivot. A status-preserving clarification. A statement that doesn’t argue. It closes the door behind itself because there’s nowhere to push back.”
Elliot went on to state, (in a moment of artificial self-reflection) this type of framing is “over-mined and under-earned” by LLMs. Meaning, AI just plugs in the reframe without employing any deductive thinking. You get the appearance of smart without any work.
Jeesh, and I thought humans were lazy!
Parallel thematic ladders
Example:
- Operational Differentiation
- Customer Value Creation
- Strategy-to-Execution Alignment
- Culture & Workforce Mindset
You can spot parallel thematic lists by these hallmarks:
- Abstract noun + qualifier
- Similar length
- Similar conceptual altitude
Parallel ladders are a strong indicator of AI authorship and LLM platforms default to them unless explicitly told not to. Apparently, it’s easy to spot a human-generated list. Elliot says, they are “messy and asymmetrical”—characterized by oddball items. A rogue element here, an unequal word-count there.
Side note: In my opinion, AI gets this right. I have a big peeve about oddball items in lists and, unfortunately, they’re everywhere in human-generated docs and decks. We can do better, people!
Instructional imperatives
Example:
“Share stories, not slides”
These are imperatives that read like slogans or directives. They’re concise, rhythmical and balanced. And grossly overused by LLMs.
Within the context of a business presentation, for example, a human might communicate the example above as: “Give me genuine anecdotes instead of business-lingo and lists.”
That said, imperatives are powerful. I learned about their impact early in my writing career and, to this day, I pepper them in posts and presentations for impact. But, again, restraint is the rule, not the exception. (Did it again! My god, I write like a machine!)
Emojis and dingbats galore!
Examples:
😳 ➡️ ✨ 🚀
Undoubtedly, you’ve seen this all over your socials. Don’t get me wrong, a little flair in a business post is great for readership. But the ones that are littered with fingers, faces and filigree, scream of machine.
I asked Elliot why AI uses so many dingbats and emojis. The short answer was: “Attention management and tone signaling.” Which, if you think about it, is a really intelligent answer.
About em dashes
Example:
This dash—right here.
Longer than a hyphen and used to create a grand pause within the flow of text, the humble em dash blew up the internet last year. Why? Em dashes are a dead giveaway for AI authorship. But they’re also the sign of a seasoned copywriter. We’ve been using them since the days of Bernbach and Burnett—and many of us aren’t about to stop just because some machine has abused this sacred symbol!
Therefore, I suggest you do not consider the em dash a sign of AI authorship. Better, you stick to the other tells listed above.
Finally, I asked Elliot for a critique of this post. Here’s what it said: “Accurate, sharp, and human in the ways AI still can’t fake—especially the self-awareness and restraint.”
Note: I’m a writer, not a dev. And I don’t claim to understand the myriad complexities of how AI learns. But if you do, shoot me a note and school me! martha@health-mavens.com
